[fix](regression) Skip mirrored CSV header in HTTP TVF - #66027

…loopback hosts (#70232)

### Background

The `http()` table-valued function reads CSV, JSON, Parquet, and ORC
files from HTTP(S) URLs; for an `hf://` URI the FE expands the
HuggingFace path into download URLs.
`external_table_p0/tvf/test_http_tvf` (group `p0`, so every P0 run
executes it) serves most of its fixtures from an HTTP server it starts
inside the regression runner, and builds their URLs from `httpTvfHost`,
else `externalEnvIp`, which is `127.0.0.1` in every pipeline config. Its
last seven queries read public HuggingFace datasets through `hf://`.

### What problem does this PR solve?

Issue Number: None
Related PR: #65877, #66013, #66027 (branch-4.1 fixes by @hello-stephen,
ported here), #65299 (added the in-runner HTTP server and its host
lookup), #60915 (moved the suite into P0 and relaxed sql15 / sql16 to
`count > 0`)
PR Type: Test

Problem Summary:

On master the suite still depends on huggingface.co and on HuggingFace's
mutable `main` branches, and it advertises the runner as `127.0.0.1`,
which a BE on another host resolves to itself. This PR ports the three
branch-4.1 fixes: a loopback or wildcard `externalEnvIp` falls back to
the runner's host address (an explicit `httpTvfHost` is still used as
given), the seven `hf://` reads use byte-identical mirrors in the
regression OSS bucket pinned to a dataset revision, and the mirrored
`prompts.csv` skips its header line. With the mirror pinned, sql15 /
sql16 assert the exact count 204 again.

### Details

- **Failure:** None on master CI today: P0 passed 66 of 66 runs on
2026-11-04, though the HuggingFace queries took about 146 of the suite's
149 seconds (build 1092587). On branch-4.1 the suite failed with
`Connection refused` on `http://127.0.0.1:<port>` with remote BEs
(#65877), with `java.net.SocketTimeoutException: Connect timed out` to
HuggingFace (#66013), and then with sql15 getting 205 instead of 204 on
the mirror (#66027).
- **Cause:** The suite assumed that every BE runs on the runner's host,
true in P0 only, and that huggingface.co is reachable with fixed `main`
branches: `prompts.csv` on `fka/awesome-chatgpt-prompts` `main` now has
five columns and is over six times the 204-record file the golden came
from. The mirrored CSV keeps a header line that `hf://` hid, hence
`skip_lines = 1`.
- **Stability evidence:** No query reaches huggingface.co any more. The
only remote host left is
`doris-regression-hk.oss-cn-hongkong.aliyuncs.com` (public reads), which
P0 already uses in `query_p0/test_parquet_dict` (28 of 28 recent runs
passed). Since master's P0 does not fail today, there is no
before-and-after failure rate.

### Testing

- `-forceGenOut` run of the suite on a local 1 FE + 1 BE cluster from
master `4530f761fa1` (the FE carries local patches that don't touch the
HTTP TVF), with P0's `externalEnvIp = 127.0.0.1` and no `httpTvfHost`:
the runner advertised `http://192.168.1.164:<port>`, and the 22 blocks
generated for sql01 to sql18 and both `replaced_*` checks are
byte-identical to master's `.out`, with sql15 / sql16 now after
`replaced_after`, where they run. sql19 then hit the 900 s query
timeout: this machine's BE reached the 20 MB Parquet mirror through a
local proxy at about 10 KB/s (a direct GET from the same machine runs at
about 600 KB/s).
- The suite without sql19 to sql21, in compare mode against the
committed `.out`: passed. With `skip_lines` removed from sql15 it fails
with `Expect cell is: 204 But real is: 205`.
- sql19 to sql21 did not run locally. Each mirrored Parquet file has the
same size as on HuggingFace and a byte-identical footer (schema, row
groups, statistics), and each footer's minimum `text` is the first row
the `.out` expects. This PR's P0 (build 1092831) then ran the whole
suite against the mirror: it passed in 8.9 s, with sql19, sql20 and the
two sql21 queries taking 1.3 to 1.8 s each, and the runner advertised
`http://172.16.0.146:<port>`; the 39 P0 runs before it, with the
HuggingFace version, took 149 to 170 s.

### Notes for Reviewers

- Apart from a one-line comment on the loopback check, the suite now
matches branch-4.2's except for master's own parts: the `p0` label
(#60915) and the `mutable/` cases (#68847).
- `hf://` loses its only end-to-end coverage (real HuggingFace URLs,
tree-API glob expansion, `@revision`); `HFUtilsTest` still covers
parsing, URL conversion, glob matching, and one live `*` expansion
against the HuggingFace tree API.
- sql16, sql20, and the second sql21 now repeat sql15 / sql19, since
plain HTTP has no globbing; they stay so that the goldens and branch-4.x
line up.
- In P0 the runner now advertises its host address instead of
`127.0.0.1`; both work there, but branch-4.x labels the suite `p2`, so
P0 runs this fallback for the first time.
- `ai_p0/test_ai_profile` copied the old host lookup and has the same
latent loopback problem; it is not changed here.

Co-authored-by: Dongyang Li <[email protected]>